Field Notes

Practical writing on AI privacy controls

Guidance for security, legal, and technology leaders bringing generative AI into regulated Canadian workflows.

AI Governance

Browser-Resident AI Privacy Controls: The Fastest Path to Secure Enterprise AI Adoption

The next breach will not begin with a dramatic intrusion. It will begin with a well-meaning employee pasting a client file, an access token, or a patient note into a generative AI window. Browser-resident AI privacy controls close that gap at the point of use, before sensitive data leaves the endpoint.

For CISOs, the appeal is architectural: no prompt proxy, no new data lake, no vendor-side inspection queue. A local pseudonymization engine can detect structured identifiers in the browser, replace them with realistic synthetic values, and preserve enough context for the AI model to remain useful. The organization reduces exposure while employees keep the productivity gains that made AI irresistible in the first place.

The best AI governance programs treat prompt security as a control plane, not a memo. Policies define what data may appear in prompts; browser controls enforce those policies in real time; dashboards report attempted disclosure as metadata. That is how regulated teams move from fear of shadow AI to measured, auditable adoption.

Compliance

AI Compliance Gap: Why Privacy Policies Fail Without Technical Enforcement

Privacy policies often say the right thing: do not share personal information, payment data, secrets, or regulated records with unauthorized AI systems. The operational problem is that employees work faster than policies can be reread.

A serious AI compliance program links policy language to controls employees cannot accidentally bypass. For Canadian organizations, that means mapping PIPEDA, Quebec Law 25, contractual confidentiality, and sector-specific obligations to concrete prompt protections: detection rules, pseudonymization, blocking thresholds, and metadata-only reporting.

  • Classify which AI workflows may include regulated data.
  • Enforce prompt DLP before text or files reach an AI platform.
  • Log events without storing raw personal or confidential data.
Security Architecture

Pseudonymization vs Redaction: The Security Difference That Keeps AI Useful

Redaction is blunt. It removes risk by removing meaning. Pseudonymization is more disciplined: it substitutes sensitive identifiers with realistic placeholders so the AI system can still reason about roles, formats, relationships, and workflow context.

In cybersecurity terms, that distinction matters. A payment dispute, SOC ticket, HR case, or support escalation can lose analytical value when every useful field becomes [REDACTED]. Format-preserving pseudonymization keeps the prompt coherent while ensuring the model does not receive the real credit card number, email address, API key, IP address, or customer identifier.

The strongest implementations keep the mapping ephemeral and local. If restoration happens inside the browser session, the AI provider sees synthetic twins, the employee sees usable output, and the enterprise avoids building another repository of sensitive prompt data.

DLP Strategy

Prompt DLP Checklist: What to Inspect Before AI Data Leaves the Browser

The most effective prompt DLP programs start with the data attackers prize and regulators scrutinize: credentials, financial identifiers, personal information, infrastructure details, and confidential business terms.

Inspection should happen before submission, not after a platform stores chat history. Security teams should combine deterministic pattern matching with managed dictionaries for company-specific terms such as project names, acquisition code words, privileged system names, and customer account labels.

  • Secrets: API keys, passwords, tokens, private keys, and access credentials.
  • PII: email addresses, phone numbers, SSNs, SINs, tax IDs, and postal identifiers.
  • Enterprise data: internal hostnames, IP ranges, financial codes, and managed keywords.
Operations

Managed Browser AI Security: Rolling Out Guardrails with GPO and MDM

Enterprise AI security succeeds when the control reaches the place where employees actually work. For many organizations, that place is the managed browser.

A browser-based AI guardrail can be distributed through familiar channels: Chrome and Edge policies, GPO, MDM, extension force-install lists, and centrally managed configuration. That lets security teams set detection rules, block high-risk submissions, and tune policies without asking every employee to become a data protection expert.

The rollout sequence should be pragmatic: monitor metadata first, tune false positives, enforce on the highest-risk categories, then expand coverage by department. Done well, managed browser security turns AI adoption into a governed workflow instead of a help desk campaign.